Search papers, labs, and topics across Lattice.
1
2
5
LLMs align even better with human preferences when trained on *less* data, revealing that preference signals are surprisingly concentrated in the initial tokens of responses.
